gcc 4.9 generates bad code for longjmp in local scope

[Stefan Hoelldampf <stefan.hoelldampf@…>]

I have successfully compiled openmodelica-1.9.2-beta1 (r24355) on Fedora 20 x86_64,
the simulation of examples works fine.

On Fedora 21 x86_64 however, the der() operator causes problems, e.g.:

1 20:47:03 Translation Error
Error occurred while flattening model Modelica.Mechanics.Translational.Examples.SignConvention

2 20:47:03 Scripting Error
/var/tmp/openmodelica-1.9.2-beta1/lib/omlibrary/Modelica 3.2.1/Mechanics/Translational.mo: 1439:7-1439:17: The der() operator is not allowed in function context (possible solutions: pass the derivative as an explicit input; use a block instead of function).

Any hints? Diffing the console outputs of both compile runs does not give a pointer.

Some differences between Fedora 20 and Fedora 21, which I have seen so far:

Fedora 20: gcc-4.8.3, java-1.7.0-openjdk
Fedora 21: gcc-4.9.2, java-1.8.0-openjdk

In SimulationRuntime/c/util/java_interface, I had to update antlr-3.1.3.jar to antlr-complete-3.5.2.jar for Java 8, but this will be another ticket...

[Martin Sjölund]

Try compiling with clang on Fedora 21. GCC tends to perform unsafe optimizations.

[Gustaf Thorslund]

I had same issue on Debian Jessie x86_64 with GCC 4.9.1 (Debian 4.9.1-19) and omc 1.9.2+dev (r24510) from subversion. A minimal test case from ​http://book.xogeny.com/behavior/equations/first_order/:

class FirstOrder
  Real x;
equation
  der(x) = 1.0 - x;
end FirstOrder;

For me changing to clang seems to fix the issue.

[Adrian Pop]

Can you check without revision r24497?
https://trac.openmodelica.org/OpenModelica/changeset/24497

I mean revert this revision and check if this problem still persists with GCC 4.9.1.
Unfortunately we don't have access to GCC 4.9.1 to check this.

[Stefan Hoelldampf <stefan.hoelldampf@…>]

On Fedora 21 with GCC, I just tried r24535 with and without r24497 applied, the problem remains in both cases. The GCC RPM system package is "gcc-4.9.2-1.fc21.x86_64".

clang will be the next try...

[Gustaf Thorslund]

Removing r24497 didn't help for me either. I've also tried earlier versions and the not so exciting result is that the first version where the bug appears when using GCC 4.9.1 on Debian testing/Jessie/8.0 is r24289. The same commit including the additional check and error message.

[Martin Sjölund]

That's weird. Because according to the GCC 4.9 tests we added, gcc-4.9 does not only fail with this message but some other ones equally surprising.

[Gustaf Thorslund]

Well, this was the only test I tried with the various versions. I suppose running more tests could result in more failures. If you like doing brain surgery with a blindfold, you might got some mud to dive into now.

[Adrian Pop]

I just tried with GCC 4.9.2 and -g instead of -O2 and all tests pass:
​https://test.openmodelica.org/hudson/job/OpenModelica_TEST_GCC_4.9/

I will play with the optimization flags a bit and see if i can find
out which one is to blame. I'm searching now for GCC 4.9.2 changes
compared to 4.8.2 so I can see if there are new optimizations included
in -O2.

[Adrian Pop]

Here is the optimization diff:

(a) gcc482_txt vs. (b) gcc492_txt

@@ -1 @@
-GCC 4.8.2 -O/O1
+GCC 4.9.2 -O/O1
           -fauto-inc-dec 
           -fcompare-elim 
           -fcprop-registers 
@@ -31 +31 @@
           -ftree-ter 
           -funit-at-a-time
           -fomit-frame-pointer (maybe)
-GCC 4.8.2 -O2
+GCC 4.9.2 -O2
           -fthread-jumps 
           -falign-functions  -falign-jumps 
           -falign-loops  -falign-labels 
@@ -39 +39 @@
           -fcrossjumping 
           -fcse-follow-jumps  -fcse-skip-blocks 
           -fdelete-null-pointer-checks 
-          -fdevirtualize 
+          -fdevirtualize -fdevirtualize-speculatively 
           -fexpensive-optimizations 
           -fgcse  -fgcse-lm  
           -fhoist-adjacent-loads 
@@ -46 +46 @@
           -finline-small-functions 
           -findirect-inlining 
           -fipa-sra 
+          -fisolate-erroneous-paths-dereference 
           -foptimize-sibling-calls 
           -fpartial-inlining 
           -fpeephole2 
-          -fregmove 
           -freorder-blocks  -freorder-functions 
           -frerun-cse-after-loop  
           -fsched-interblock  -fsched-spec 

[Adrian Pop]

Even with -O1 it has issues. It seems that some of the optimizations in -O1 are bad for us.

[Stefan Hoelldampf <stefan.hoelldampf@…>]

Passing
CFLAGS="-fno-tree-bit-ccp"
to the configure script solves the problem for r24582 on Fedora 21 (gcc-4.9.2-1.fc21.x86_64).

I also tried to extend the existing flag detection for GCC 4.8 in configure.in, but the attached patch does not fix the problem. Perhaps the environment variable CFLAGS is used somewhere else not covered by the flag detection magic...

[Stefan Hoelldampf <stefan.hoelldampf@…>]

openmodelica-r24582-gcc4.9-flags.patch (does not fix the problem!)

[Martin Sjölund]

If you pass CFLAGS="-fno-tree-bit-cpp", you will automatically use -O0, right? So it probably means it is not the interesting flag.

[Martin Sjölund]

I will try:

echo "Finding result" > log
for flag in \
 -fauto-inc-dec -fcompare-elim -fcprop-registers -fdce -fdefer-pop -fdelayed-branch -fdse -fguess-branch-probability -fif-conversion2 -fif-conversion -fipa-pure-const -fipa-profile -fipa-reference -fmerge-constants -fsplit-wide-types -ftree-bit-ccp -ftree-builtin-call-dce -ftree-ccp -ftree-ch -ftree-copyrename -ftree-dce -ftree-dominator-opts -ftree-dse -ftree-forwprop -ftree-fre -ftree-phiprop -ftree-slsr -ftree-sra -ftree-pta -ftree-ter -funit-at-a-time \
 -fthread-jumps -falign-functions -falign-jumps -falign-loops -falign-labels -fcaller-saves -fcrossjumping -fcse-follow-jumps -fcse-skip-blocks -fdelete-null-pointer-checks -fdevirtualize -fdevirtualize-speculatively -fexpensive-optimizations -fgcse -fgcse-lm -fhoist-adjacent-loads -finline-small-functions -findirect-inlining -fipa-sra -fisolate-erroneous-paths-dereference -foptimize-sibling-calls -fpartial-inlining -fpeephole2 -freorder-blocks -freorder-functions -frerun-cse-after-loop -fsched-interblock -fsched-spec -fschedule-insns -fschedule-insns2 -fstrict-aliasing -fstrict-overflow -ftree-switch-conversion -ftree-tail-merge -ftree-pre -ftree-vrp \
 -finline-functions -funswitch-loops -fpredictive-commoning -fgcse-after-reload -ftree-loop-vectorize -ftree-slp-vectorize -fvect-cost-model -ftree-partial-pre -fipa-cp-clone
do
 echo $flag
 ./configure "CFLAGS=$flag" 'CC=gcc-4.9' 'CXX=g++-4.9' 'OMPCC=gcc-4.9 -fopenmp' '--disable-modelica3d' '--with-omniORB' '--with-omc=/usr/bin/omc'
 make -j40 clean
 if make -j40 sanity-check; then
 echo "$flag OK" >> log
 else
 echo "$flag FAILED" >> log
 fi
done

[Martin Sjölund]

All of the flags succeeded. So it is some combination of flags? :(

[Gustaf Thorslund]

I did some less successful flag testing last night. To get the difference between -O0 (the one working) and -O1 (the first one failing) I used:

  diff -u <(gcc -Q --help=optimizers -O0) <(gcc -Q --help=optimizers -O1) |\
  grep "^+ " | cut -d" " -f3

-fbranch-count-reg
-fcombine-stack-adjustments
-fcompare-elim
-fcprop-registers
-fdefer-pop
-fforward-propagate
-fguess-branch-probability
-fif-conversion
-fif-conversion2
-finline-functions-called-once
-fipa-profile
-fipa-pure-const
-fipa-reference
-fmerge-constants
-fmove-loop-invariants
-fshrink-wrap
-fsplit-wide-types
-ftree-bit-ccp
-ftree-ccp
-ftree-ch
-ftree-copy-prop
-ftree-copyrename
-ftree-dce
-ftree-dominator-opts
-ftree-dse
-ftree-fre
-ftree-pta
-ftree-sink
-ftree-slsr
-ftree-sra
-ftree-ter

I grouped them into three groups and compiled with CFLAGS="-O0 $(cat group-$i)", but none of them failed. Then I tried adding all flags, so compiling with:

CFLAGS="-O0 -fbranch-count-reg -fcombine-stack-adjustments -fcompare-elim -fcprop-registers -fdefer-pop -fforward-propagate -fguess-branch-probability -fif-conversion -fif-conversion2 -finline-functions-called-once -fipa-profile -fipa-pure-const -fipa-reference -fmerge-constants -fmove-loop-invariants -fshrink-wrap -fsplit-wide-types -ftree-bit-ccp -ftree-ccp -ftree-ch -ftree-copy-prop -ftree-copyrename -ftree-dce -ftree-dominator-opts -ftree-dse -ftree-fre -ftree-pta -ftree-sink -ftree-slsr -ftree-sra -ftree-ter "

But that one didn't fail the test either. Maybe it would be worth removing one flag at the time from -O1 or -O2, instead of just trying all flags one at the time.

[Martin Sjölund]

I have a possible fix in the works (replacing MMC_THROW_INTERNAL longjmp'ing to the local function with goto). It seemed to work for the simple testcase; running the testsuite now.

[Martin Sjölund]

Try r24585

[Adrian Pop]

If not, this will definitely fix it properly (and is just a bit slower), but I will make a version that has some ifdefs based on compiler versions:

Compiler/Template/CodegenC.tpl

@@ -10043 +10043 @@
         then 'tmpMeta[<%System.tmpTickIndex(1)%>]'
       else
         let newVarIx = 'tmp<%System.tmpTick()%>'
-        let &varDecls += '<%ty%> <%newVarIx%>;<%\n%>'
+        let &varDecls += 'volatile <%ty%> <%newVarIx%>;<%\n%>'
         newVarIx
   newVar
 end tempDecl;

[Stefan Hoelldampf <stefan.hoelldampf@…>]

Thanks, r24585 works on Fedora 21 (gcc-4.9.2-1.fc21.x86_64) without additional patches and without CFLAGS being set.

I did some more testing regarding CFLAGS and r24584, the following "no" flags also fix the problem:

CFLAGS="-g -O2 -fno-if-conversion -fno-if-conversion2"
("-g -O2" mimics the default CFLAGS)

[Martin Sjölund]

Thanks everyone who helped try to find out what went wrong. We have a bit smarter code generation now. We could get some additional performance improvements with a better IR (fewer setjmp/longjmp). But at least the code is more correct now than before.